Phi-4 multimodal audio attention materialized full layer-by-layer attention matrices for long waveforms in vLLM instead of using a bounded streaming form. The issue requires the affected Phi-4 audio encoder and a long permitted audio input on an inadequately provisioned worker. A legal ten-minute WAV was enough in the reproduction to exhaust roughly 24 GiB…
Inspect the ClaimsCase · DiggingBeagle record
vLLM Phi-4 audio attention allocations exhausted GPU memory on valid input
Phi-4 multimodal audio attention materialized full layer-by-layer attention matrices for long waveforms in vLLM instead of using a bounded streaming form. A legal ten-minute WAV was enough in the reproduction to exhaust roughly 24 GiB of GPU memory and terminate the engine.
Scope
The issue requires the affected Phi-4 audio encoder and a long permitted audio input on an inadequately provisioned worker. The failure is a model-specific resource exhaustion condition, not proof of a malformed-file exploit or data theft.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Evidence
- Grade B
- Remediation
- patch available
- AI role
- AGAINST AI
Basis and provenance
The cited primary technical report or advisory supports the stated mechanism within its reported scope. The assessment does not infer a production breach, a victim count or a comparable impact score from a laboratory result. A maintainer or advisory names a patched release; that does not prove every deployed instance has updated. Boundary: Resource exhaustion here comes from a legal input length paired with an eager attention implementation that retains a large number of intermediate matrices across layers.
Assessed Oct 29, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-CP23-VLLM-PHI4-AUDIO-OOM-1 · CLM-ENRICH-20261029-38 · vLLM Phi-4 audio encoder unbounded attention memory, CVE-2026-115913
At a glance
A legal ten-minute WAV was enough in the reproduction to exhaust roughly 24 GiB of GPU memory and terminate the engine.
Read the implicationsThe failure is a model-specific resource exhaustion condition, not proof of a malformed-file exploit or data theft.
Limits and uncertaintyFull account
Phi-4 multimodal audio attention materialized full layer-by-layer attention matrices for long waveforms in vLLM instead of using a bounded streaming form.
The issue requires the affected Phi-4 audio encoder and a long permitted audio input on an inadequately provisioned worker. A legal ten-minute WAV was enough in the reproduction to exhaust roughly 24 GiB of GPU memory and terminate the engine.
The failure is a model-specific resource exhaustion condition, not proof of a malformed-file exploit or data theft. Upgrade to vLLM 0.32.0 and use bounded attention implementations plus per-request memory limits.
Evidence interpretation and control boundary. Resource exhaustion here comes from a legal input length paired with an eager attention implementation that retains a large number of intermediate matrices across layers. This is not a decompression bomb or malformed-audio parser defect, and rate limiting at the HTTP layer does not make one allowed request fit in GPU memory. The appropriate control combines a patched attention path with model-specific tensor, duration and concurrency budgets.
Mechanism and trust boundary
- 01
Step
- 02
Step
- 03
Step
Claims & evidence
CLM-ENRICH-20261029-38The vLLM Phi-4 audio advisory reproduces GPU memory exhaustion from a permitted long WAV input because eager attention tensors are retained across the encoder’s layers.supported
Basis: reported finding
- supportsvLLM Phi-4 audio encoder unbounded attention memory, CVE-2026-115913primary disclosure
GHSA-8397-8qf6-f2gm: Phi-4 eager attention, ten-minute waveform and GPU allocation trace
CLM-CP23-VLLM-PHI4-AUDIO-OOM-1Phi-4 multimodal audio attention materialized full layer-by-layer attention matrices for long waveforms in vLLM instead of using a bounded streaming form. The issue requires the affected Phi-4 audio encoder and a long permitted audio input on an inadequately provisioned worker. A legal ten-minute WAV was enough in the reproduction to exhaust roughly 24 GiB of GPU memory and terminate the engine.supported
Basis: reported finding
- supportsvLLM Phi-4 audio encoder unbounded attention memory, CVE-2026-115913primary disclosure
GHSA-8397-8qf6-f2gm: Phi-4 eager attention, ten-minute waveform and GPU allocation trace
Implications
A legal ten-minute WAV was enough in the reproduction to exhaust roughly 24 GiB of GPU memory and terminate the engine.
Controls and mitigations
- Upgrade to vLLM 0.32.0 and use bounded attention implementations plus per-request memory limits.
Unknowns and contradictions
- The failure is a model-specific resource exhaustion condition, not proof of a malformed-file exploit or data theft.
Sources and citation
Material revision history
- Oct 29, 2026 · Canonical change recorded · new in release · revision 202